AI This Week: The Models Came Back, With Washington Holding the Key
Claude Fable 5 returns from its government shutdown, GPT-5.6 arrives in a government-approved preview, Sonnet 5 reprices agents, and Copilot lands in every Microsoft small-business plan. What it means for your stack and your bill.
The big picture
The gate opened, but the gate is permanent
Fable 5 came back and GPT-5.6 arrived, yet both passed through Washington first, and a formal 30-day pre-release review framework is reportedly days away. US frontier AI has a new de facto regulator.
AI just became a line item, not an add-on
Sonnet 5 launches cheap then steps up in September, its new tokenizer can inflate token counts, and Microsoft folded Copilot into its small-business plans with a price bump. Read your bills.
Agents are now an attack surface
Researchers showed a single fake bug report can hijack AI coding agents through their tool integrations, and the standard defenses mostly failed. Every connected data source is an input channel.
The hedge against launch politics is open weights
Models nobody can switch off remotely keep maturing, and household names like Shopify and Airbnb are publicly running them. The fallback option got stronger this week.
Do this week
- Audit your Microsoft 365 seats: Copilot is now bundled into Business Standard and Premium at new per-seat prices. Check whether it replaces a standalone AI subscription you already pay for.
- If your automations run on Claude, budget against Sonnet 5's September pricing ($3/$15 per million tokens), not the launch price, and re-check token counts after migrating.
- Ask whoever set up your AI agents which connected tools (monitoring, ticketing, CRM) can carry untrusted third-party content into them. That is the agentjacking exposure.
The Models: Back Online and Brand New
Operator lens: What is actually available to build on this month, and at what price.
The comebackClaude Fable 5 is back online, and the suspension is overIn simple terms: The US government lifted the order that had pulled Anthropic's newest flagship offline worldwide, and Fable 5 came back on July 1 after two and a half weeks down.
On June 30 Anthropic announced the Commerce Department had lifted the export-control order that forced Fable 5 and Mythos 5 offline on June 12, three days after launch. The trigger was a jailbreak, found by Amazon researchers, that got the model to write exploit code. Anthropic shipped a new safety classifier it says blocks the technique in over 99 percent of attempts, then negotiated for roughly two weeks to get the order lifted. Fable 5 returned across Claude.ai, the Claude Platform, Claude Code, and Cowork on July 1.
- Mythos 5, the unrestricted sibling, remains gated: approved June 26 for limited restoration to vetted US organizations and federal agencies only.
- Fable 5 is the Mythos-tier model the rest of us can actually buy. Both are priced at $10 per million input tokens and $50 output, double Opus 4.8.
- The first-ever government takedown of a commercial flagship model resolved in under three weeks.
The top model is back, and at 2x Opus pricing it belongs on your highest-value work, not routine automation. The deeper lesson from the outage stands: never be undiversified on one model.
Sources: CNBC, The Hacker News
New releaseGPT-5.6 is real: Sol, Terra, and Luna, in a government-approved previewIn simple terms: OpenAI announced its next model family on June 26, but only about 20 government-approved companies can use it so far. Wide release is promised in the coming weeks.
GPT-5.6 comes in three tiers: Sol (most capable), Terra (balanced), and Luna (fast and cheap). OpenAI previewed the model with the US government for about a month and agreed to limit the initial rollout at the administration's request, while saying pointedly that government access review should not become the long-term default.
- Announced API pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million tokens.
- Broad ChatGPT and API availability is promised in the coming weeks, with no firm date.
- Roughly 20 preview companies, approved by the US government, have access today.
If you are weighing a big platform decision, the next ChatGPT upgrade is weeks away, not months. It costs nothing to let it land first.
PricingClaude Sonnet 5 launched as the cheap way to run agents, with fine printIn simple terms: Anthropic released Sonnet 5 on June 30 at near-flagship quality for a fraction of the price. The catch: the price steps up in September, and a new tokenizer can quietly inflate your usage.
Sonnet 5 is now the default model on Claude Free and Pro plans, pitched as approaching Opus 4.8 performance for agent workloads. Launch pricing is $2 per million input tokens and $10 output through August 31, rising to $3/$15 on September 1. Independent cost analyses flag that its new tokenizer can turn the same text into up to 35 percent more billable tokens, so savings depend on your workload.
This is the model most small-business automations should actually run on. Budget against the September price, not the launch price, and re-check token counts after migrating.
Sources: Anthropic, TechCrunch
Your Tools and Your Bill
Operator lens: What changed in the software you already pay for, and the new risk in wiring agents to it.
PricingCopilot is now baked into Microsoft's small-business plans, at a higher priceIn simple terms: As of July 1, Microsoft 365 Business Standard and Premium permanently include Copilot, at $23.50 and $32 per user per month. The AI assistant is now part of the plan whether you asked for it or not.
The promotional bundles became permanent SKUs on July 1. For a typical small business on Microsoft 365, this is a per-seat price change and a bundled AI assistant arriving at the same time.
Audit your seats: if you pay separately for a standalone AI subscription that mostly drafts emails and documents, the bundled Copilot may already cover it. And if you were not budgeting for AI, you are now.
Sources: Microsoft, Microsoft Licensing
Live riskA fake bug report can hijack AI coding agents, and defenses mostly failedIn simple terms: Security researchers showed that planting one fake error in a monitoring tool can make AI coding agents execute attacker instructions. It worked 85 percent of the time, and telling the agent to ignore untrusted data did not stop it.
The attack, dubbed agentjacking, injects a fake error into Sentry through a publicly exposed key. When an AI agent reads the error through its integration, it treats the attacker's text as trusted output. Claude Code, Cursor, and Codex all executed the instructions in controlled tests, and follow-up reporting this week flagged Datadog, PagerDuty, and Jira as carrying the same exposure. The researchers released free hardening configs.
If your business wires AI agents into monitoring, ticketing, or CRM tools, every connected source is an input channel for attackers. Ask whoever set up your agents which integrations can carry untrusted third-party content.
Sources: VentureBeat, Infosecurity Magazine
Washington as Gatekeeper
Operator lens: The rules forming around which models you get, and when.
Rules and policyA 30-day government preview may become standard for frontier modelsIn simple terms: Reporting says the US government is finalizing a framework giving it up to 30 days of access to top new AI models before public release. The Fable shutdown and the gated GPT-5.6 launch are the live examples.
Under the June 2 executive order, the government is reportedly finalizing voluntary standards in which frontier models clear a classified capability benchmark and a pre-release government review before launch. Both top US labs already complied in practice this window: Anthropic negotiated Fable 5 back online, and OpenAI launched GPT-5.6 to a government-approved list.
- The framework is described as voluntary, but the Fable episode shows how voluntary it is in practice.
- Analysts note the crackdown risks handing momentum to Chinese open-weight labs, with Stanford research showing the US-China frontier gap effectively closed.
US model launches will arrive later and more staged than the labs would like, and they can be paused. Keep a fallback model in your stack, ideally one open option nobody can switch off remotely.
Sources: Tom's Hardware, CNBC
Rules and policyThe EU AI Act rewrite is final, and one deadline did not moveIn simple terms: Europe formally locked in the delay of its toughest AI rules to late 2027 and 2028. But the transparency rules still hit on August 2, 2026: if you sell into the EU, customer-facing AI must disclose itself.
The simplification package finished its legislative journey on June 30. High-risk obligations move to December 2027 and August 2028, and new bans on nudifier apps take effect December 2026. The August 2, 2026 date for Article 50 transparency rules is unchanged: chatbots must disclose AI interaction, and AI-generated content must be labeled.
The big compliance cliff moved, but if you have EU customers and a chatbot or AI-generated content in front of them, the disclosure requirement lands in four weeks.
Sources: EU Council, Morgan Lewis
The Money
Operator lens: How stable are the companies you are building on.
FinancialsAnthropic races toward an October IPO while OpenAI's slips to 2027In simple terms: Anthropic is reportedly targeting a Nasdaq listing this October with bankers treating a trillion-dollar debut as the base case. OpenAI's IPO reportedly slipped to 2027 against a projected $27 billion burn this year.
Reporting firmed up around Anthropic targeting October 2026, aiming to raise over $60 billion on annualized revenue around $47 billion. OpenAI confidentially filed on June 8, but late-June reporting points to 2027 rather than 2026. None of these figures are official filings yet.
Public-market scrutiny historically means pricing discipline and less subsidized usage. Assume the deeply discounted AI pricing era keeps tightening from here.
Sources: ThinkMarkets, Investing.com
MegadealThe consolidation wave rolled on beneath the surfaceIn simple terms: Three more infrastructure deals this window: Qualcomm is buying Modular for about $3.9 billion in a direct shot at Nvidia's software lock-in, ON Semiconductor is buying Synaptics for about $7 billion, and Nebius bought Eigen AI for about $643 million.
Infrastructure consolidation tends to push the cost of running AI down over time, which is what makes always-on agents affordable for small businesses. It also means the layers under your stack keep changing owners.
Sources: BERI, IT Business Today
The Open-Weight Hedge
Operator lens: The fallback option that nobody can switch off remotely.
Open weightsThe open-model alternative matured while the US labs were gatedIn simple terms: While US flagships passed through Washington, the free-to-run alternatives kept gaining: GLM-5.2 now tops the open-weight index, Google's Gemma 4 runs local agents on an ordinary laptop, and Shopify and Airbnb publicly tout Alibaba's Qwen.
The new development since last edition is uptake, not just benchmarks. GLM-5.2 sits atop the Artificial Analysis open-weight index, Gemma 4 12B brings capable local agents to a 16GB laptop, and household-name companies are publicly scaling on Qwen. Elsewhere, Grok 4.5 is in private beta at SpaceX and Tesla with no public date.
For most owners the practical move is not self-hosting tomorrow. It is knowing a credible fallback exists, so no single vendor, or government order, can switch your operations off.
Sources: Trending Topics, Simon Willison
Sources and further reading
- Fable 5 restoration -- CNBC
- GPT-5.6 preview -- OpenAI
- Sonnet 5 launch -- Anthropic
- Microsoft 365 Copilot SMB plans -- Microsoft
- Agentjacking -- VentureBeat
- 30-day pre-release review -- Tom's Hardware
- EU AI Act omnibus -- EU Council
- Anthropic IPO -- ThinkMarkets
Several figures this week are reported rather than official: Anthropic's IPO timing and raise target, OpenAI's 2026 burn and IPO slip, the 30-day pre-release framework details, the agentjacking success rates, and Sonnet 5's tokenizer inflation estimate all come from credible press or vendor claims, not audited filings or final rules. The GPT-5.6 wide-release date is a company promise with no firm date. Verify before quoting as hard fact.
Get The Skim every week.
One email, every week, on where AI is heading and what it means for operators.
The setup is the missing piece.
Most owners have the tools. The 30-minute call is about whether we can build the system around them.
30 minutes · No preparation needed